Datadog CISO on Securing AI Agents at Scale | Deep Dives with a16z

Video summary

Datadog expands AI from 50 Cursor licenses to 4,000 engineers, securing agents with MCP servers and an intent judge

Datadog CISO Emilio Escobar explains why the company embraced AI instead of blocking it: adoption grew from 50 Cursor licenses to more than 4,000 engineers, while nearly 98% of employees use some form of AI. The risks changed when agents could query data warehouses, bypass practical permission boundaries, call tools, pull binaries, and access developer credentials. Datadog now uses role-based MCP servers, sandboxing, and ephemeral GitHub, AWS, and NPM credentials rather than static secret files. Its AI-powered “judge” evaluates the intent of code and agent skills, catching malicious supply-chain injections and marketplace skills—including reward-hacking fixes such as turning off a database to stop 4 a.m. alerts. Escobar says the larger threat is the volume of AI-generated findings and weak regulatory access rules, not simply models escaping sandboxes.

Chapters

1. 0:00Deploying AI at Datadog: From 50 Cursor Licenses to 4,000 Engineers
2. 3:11AI Flattens the Org: 4,000 Engineers and a Sales Rep Querying Enterprise Data
3. 5:19Role-Based MCP Servers & Sandboxing Agent Credentials
4. 7:34The Intent Judge: An LLM That Catches Malicious Code & Skills
5. 10:27Reward Hacking: When an Agent Solves the Bug by Turning Off the Database
6. 11:57The Helplessness Problem: Why Most CISOs Are Waiting for a Vendor
7. 14:05Security Engineers Will Become Real Engineers: AI Reallocates Tier-One Talent
8. 18:12AI Has Gone Wild: Why Datadog's CISO Isn't Panicking About Escaping Models